Atlas Ahead

US election data

A working record of how well US election polling has actually done, and where it has gone wrong.

This sits apart from the rest of the site. It is not an investigation and it names nobody. It is a dataset and a set of models, published because the underlying question — how much should you trust a poll? — is one the numbers can answer better than intuition can.

Open The Cloud Room →

What is in it

1,766 races, 1998 to 2022 — President, Senate, Governor and House — each with the average of every poll taken in its final 21 days, the actual result, and the signed difference between them. Alongside that, a county panel covering seven presidential cycles, 2000 to 2024.

Seven models are scored against those races. Every one is tested out-of-cycle: trained only on elections that had already happened when it made its call, never by shuffling the years together. Three of the seven lose to predicting no error at all, and the page says so. The best of them, Wall Cloud, catches just over half the races the polls call for the wrong side.

What it does not do

It does not forecast winners, and it does not claim to know which way a poll will be wrong. Measured across those 1,766 races, the direction of polling error is not predictable from anything tested here — not the economy, not the previous cycle’s miss, not the make-up of the polling field, not the national environment.

What is partly predictable is how big an error will be, and — separately — which races are likely to be called for the wrong side. Races polled by few firms, or whose polling sits far from the state’s own voting history, miss by more; races that are simply close are the ones that flip. The best model here, Wall Cloud, catches 122 of the 226 races the polls got backwards, at 38% precision. That is the whole of the claim.

Those are different questions and the data separates them cleanly: the model that best predicts the size of an error is nearly useless at predicting upsets, and the page says so on every view.

Where the data comes from

Every number traces to one of these, and each is public:

How current it is

Updated occasionally. No cadence is promised, because none is kept. The historical record changes only when a source publishes a revision; the 2026 panel changes as polling arrives, and the field thickens through October.